AI systems · research & engineering
Most agents look capable when everything goes right. I care about the harder cases: thousands of available skills, long-running tool use, partial failures, and tight context or latency budgets. My work connects dependency-aware retrieval, execution validation, and failure recovery—then turns those ideas into open systems others can inspect, test, and extend.
I am an AI systems researcher and software engineer. The thread through my work—from scientific visualization and distributed tracing to recommendation and agent infrastructure—is a desire to make complex technology useful, inspectable, and dependable.
The throughline
I enjoy the unglamorous middle between a promising prototype and a dependable system: evaluation, interfaces, observability, latency, failure handling, and the thousand small decisions that make software hold together. That mindset now shapes my work on agent skill retrieval, reliable execution, and AI-native research workflows.
What I’m pursuing
I’m currently a Software Engineer at TikTok in Seattle, working on Shop Ads ranking, creative intelligence, and delivery systems. Production engineering keeps my research honest: useful ideas still have to survive scale, latency, noisy inputs, incomplete observability, and real users.
In parallel, I collaborate with LAIR, advised by Prof. Lichao Sun, and contribute to Dr. Claw. I hold an M.S.E. from the University of Pennsylvania and a B.E. from Northeastern University.
01 · Retrieve
Graph-of-Skills, accepted to EMNLP 2026 (Main Conference), makes retrieval dependency-aware rather than purely semantic. On SkillsBench, its strongest setting improved reward by 25.6% (+7.0 absolute points) while reducing total tokens by 56.7% against full skill loading.
02 · Execute & recover
COMFYCLAW, accepted to ECCV 2026 Workshop MUCG, combines typed workflow editing, stage-aware tools, rollback, visual verification, and skill evolution to make executable image-generation workflows more reliable.
03 · Build in the open
As a core contributor to Dr. Claw, accepted to EMNLP 2026 (System Demonstrations), I work across provider integrations, multi-session workflows, desktop distribution, project isolation, and the reliability details behind long-running research tasks.
04 · Test against reality
Work across TikTok, Amazon, and JD Technology taught me to treat latency, observability, experimentation, and failure handling as design inputs—not cleanup after a prototype succeeds.
Ordered by how closely each one sits to my current focus rather than by date. The first three follow the same retrieve–execute–recover loop as the section above; the rest are the multimodal and scientific systems work that led there.
* Equal contribution (co-first author). † Core contributor, as designated in the cited paper.
Modern LLM agents increasingly rely on reusable skills, and as they interact with personal applications, web browsers, and other interfaces, skill libraries can scale to thousands of skills. Scaling to larger skill sets introduces two key…
Agents are increasingly used to control executable workflows rather than only answer questions. This makes reliability depend on both execution-time scaffolding and the ability to reuse experience from past runs. We study this problem in…
Command-line coding agents (e.g., Claude Code, Gemini CLI) can already read and write files and sustain long sessions, yet end-to-end research still fragments across chat tools, IDEs, terminals, and writing environments, and the decisions…
3D vision has rapidly evolved, driven by increasingly diverse data representations, learning paradigms, and modeling strategies. Yet the field remains fragmented across representations and benchmarks, making it difficult to develop unified…
Video generation has advanced rapidly, yet visual content alone is insufficient for many realistic and engaging media experiences: sound is central to immersion, semantic coherence, and perceived causality. Growing demand for synchronized…
TIMEDB is an online database for analyzing human tumor immune microenvironment cell-type composition from bulk expression profiles. It curates expression and composition profiles with clinical information for 39,706 samples from 546…
Six selected here. The complete, continuously updated record lives on Google Scholar and DBLP.
I prefer research that leaves behind something people can run. Alongside Dr. Claw above, this is the evaluation work I contribute to with the LAIR community.
A research-oriented training path grounded in both systems engineering and applied AI.
Education
Education
I like roles where modeling, infrastructure, and product rigor meet. Across these teams, I have learned how ranking, observability, platform engineering, and backend reliability behave under real constraints.
TikTok (ByteDance Inc.)
Commerce Ads Team
Work on ranking, creative intelligence, and delivery tooling for Shop Ads. Joined as an SDE Intern, converted to a full-time Software Engineer within 12 weeks, and returned full-time after graduation.
Amazon
Global-Mile Team
Focused on distributed observability infrastructure, building tracing and telemetry tools that worked across microservices and serverless environments without invasive code changes.
JD Technology
Algorithm Tools Team
Worked on platform engineering for internal AI tooling, with an emphasis on automation, modular system design, and cloud-native delivery workflows.
Competition projects where I led the build, from spatial interaction on Apple platforms to recommendation-driven web products.
Solo Developer
Winner · Swift Student Challenge 2026
A touchless cooking assistant for iPad that removes the “digital friction” of swiping recipes with messy hands, using on-device computer vision so the device fades into the background.
Team Leader
2nd Prize · Mobile Application Innovation Contest 2022
An AR social app, co-hosted by Apple and Zhejiang University, that lets users anchor and share geo-tagged notes in physical space and earn tokens through daily check-ins.
Team Leader
1st Prize · ByteDance Youth Camp 2022
A rebuilt developer community web app with a personalized “For You” feed, post interactions, and responsive design.
I keep a long-running technical knowledge base alongside my research and engineering work. The goal is not to collect definitions, but to connect first principles, problem-solving methods, and production tradeoffs.
Model · Explore · Optimize · Verify
Data structures, search, dynamic programming, optimization, correctness, and interview problem solving.
Open algorithm map →02Foundation · Model · System · Product
Deep learning, language models, inference, agents, recommendation, search, and evaluation.
Open AI map →03Runtime · Network · Design · Distributed
Operating systems, computer networks, OOD, design patterns, and scalable system design.
Open CS map →04Build · Store · Integrate · Operate
Languages, data systems, middleware, delivery, observability, testing, and production practice.
Open engineering map →05Diagnose · Drill · Simulate · Review
A personal TOEFL and GRE study system built around skills, evidence, deliberate practice, and review.
Open study map →Directory pages build the map. Topic pages explain one complete idea. Case studies connect concepts under realistic constraints. When the same subject appears in multiple areas, I prefer cross-links over duplicated definitions.
My toolkit is organized around the systems I build, not a flat inventory of technologies. These four working layers connect research prototypes to observable, production-facing products.
Model, retrieve, evaluate, and operate tool-using systems under real context and latency constraints.
Build typed services, ranking paths, streaming jobs, and the interfaces that keep them evolvable.
Design storage and delivery foundations with explicit tradeoffs around scale, failure, and observability.
Turn systems into usable research workspaces, web products, visual tools, and spatial experiences.
This is a curated working set rather than a proficiency checklist. I choose tools by the constraints of the system and keep the deeper implementation notes in the AI, CS, and software engineering knowledge maps.
Two pages
Education, agent-infrastructure research and the four engineering roles behind it on the first page; publications, honors, and research impact on the second. The same record as the sections above, in the form people ask for it. Typeset from source, so it is regenerated rather than edited by hand.
PDF · 2 pages · Updated Aug 2026
I’m always glad to meet researchers and engineers thinking seriously about agent reliability, skill retrieval, research automation, or AI systems at production scale. If you are building something where the demo is the easy part, we will probably have a useful conversation.
